On Wed, Oct 16, 2013 at 10:54:06PM +0100, Richard Curnow wrote: > Going back many years, I don't recall much rigour going into the design > of the "number of runs" logic. The code arose from a heuristic that > seemed to work well enough; I didn't find anything in the literature > that led to it being replaced with something else.
>From what I remember it seemed to be very well tuned. When I was trying to tweak the table of critical runs, any change seemed to make the results worse. The only interesting change I was able to do in this area was doubling the number of samples used in the runs test, i.e. do the regression with n samples, but count the runs also in the tail extending to the 2*nth sample. That seems to slightly improve the overall response. > If someone can find some applicable "number of runs" theory that could > apply, it would be worth looking into. I'd expect it should be possible to somehow include the sample weights in the test. I've tried few naive approaches, but failed miserably. What I'd like to try sometimes is a hybrid approach, one part looking for a long term optimal number of samples and the other part temporarily decreasing the number when an exceptional situation arises (e.g. rapid changes in the temperature). -- Miroslav Lichvar -- To unsubscribe email chrony-dev-requ...@chrony.tuxfamily.org with "unsubscribe" in the subject. For help email chrony-dev-requ...@chrony.tuxfamily.org with "help" in the subject. Trouble? Email listmas...@chrony.tuxfamily.org.
